OpenResearcher: Open-Source Deep Research Agent That Outperforms GPT-4.1 and Claude

TIGER-Lab at the University of Waterloo has released OpenResearcher, a fully open agentic large language model (30B-A3B) designed for long-horizon deep research scenarios. The model achieves 54.8% accuracy on BrowseComp-Plus, surpassing GPT-4.1, Claude Opus 4, Gemini 2.5 Pro, DeepSeek-R1, and Tongyi-DeepResearch.

What makes OpenResearcher unique is its fully offline training pipeline. It synthesizes 100+ turn deep-research trajectories without any search or scrape APIs, rate limits, or nondeterminism. The system uses a local retriever and a 10-trillion-token corpus to generate long-horizon tool-use traces, making the training process completely reproducible.

The team has fully open-sourced everything: code, search engine, corpus recipe, 96K training trajectories, evaluation logs, trained models, and a live demo. This is the most comprehensive open release for a deep research agent to date.

OpenResearcher represents a significant step toward democratizing deep research capabilities. While proprietary systems like Gemini Deep Research and Claude Research require API access and incur ongoing costs, OpenResearcher can be deployed entirely on-premises with no external dependencies.

GitHub: https://github.com/TIGER-AI-Lab/OpenResearcher
Paper: https://huggingface.co/papers/2603.20278